This issue tracker has been migrated to GitHub, and is currently read-only.
For more information, see the GitHub FAQs in the Python's Developer Guide.

Title: Add gdb7 hooks to make it easier to debug Python
Type: Stage:
Components: Demos and Tools Versions: Python 2.7
Status: closed Resolution: accepted
Dependencies: Superseder:
Assigned To: loewis Nosy List: benjamin.peterson, davidfraser, dmalcolm, doko, gregory.p.smith, jcea, jyasskin, loewis, mark.dickinson, scott.tsai
Priority: normal Keywords: patch

Created on 2010-02-28 23:40 by dmalcolm, last changed 2022-04-11 14:56 by admin. This issue is now closed.

File name Uploaded Description Edit
add-gdb7-python-hooks-to-trunk.patch dmalcolm, 2010-02-28 23:43 Patch against trunk (r78517)
add-gdb7-python-hooks-to-trunk-v2.patch dmalcolm, 2010-03-08 20:20
diff-of-gdb7-hooks-v2-relative-to-v1.diff dmalcolm, 2010-03-08 20:23 What's changed in v2 compared to v1, for reference
add-gdb7-python-hooks-to-trunk-v3.patch dmalcolm, 2010-03-11 21:11 Version 3 of patch, against trunk (r78840)
diff-of-gdb7-hooks-v3-relative-to-v2.diff dmalcolm, 2010-03-11 21:12 What's changed in v3 compared to v2, for reference
add-gdb7-python-hooks-to-trunk-v4.patch dmalcolm, 2010-03-25 16:53 Version 4 of patch, against trunk (79422)
diff-of-gdb7-hooks-v4-relative-to-v3.diff dmalcolm, 2010-03-25 16:55 What's changed in v4 compared to v3, for reference
add-gdb7-python-hooks-to-trunk-v5.patch dmalcolm, 2010-04-01 02:34 Version 5 of patch, against trunk (r79517)
diff-of-gdb7-hooks-v5-relative-to-v4.diff dmalcolm, 2010-04-01 02:37 What's changed in v5 compared to v4, for reference
Messages (22)
msg100226 - (view) Author: Dave Malcolm (dmalcolm) (Python committer) Date: 2010-02-28 23:40
gdb 7 can be extended with Python code, allowing the writing of domain-specific pretty-printers and commands.  

I've been working on gdb 7 hooks to make it easier to debug python itself, as mentioned here:

I'm attaching a patch for merger with "trunk".  My hope is to be the maintainer of this work, although I do not yet have SVN commit rights (see )

The code is fully compatible with the existing "Misc/gdbinit" macros - you can freely use both as needed.

= Two versions of Python =
This code is intended to run inside gdb.  There are potentially two Python versions involved here: that of the "inferior process" (the one being debugged), and that of the debugger.

As I understand things, gdb only supports embedding Python 2 at the moment.  This code is thus targeting that version of Python.

So far, I've attempted to keep this code so that it will run when the "inferior process" is either Python 2 or Python 3.  I could vary this code in the py3k branch if desired.  The code would then track the "inferior process" version of Python, so that code to debug Python 3 would live in the py3k branch.  That code would still be written for Python 2, though.

= Makefile changes =

The Makefile installs the gdb hooks to relative to the built python (even if you're not using them), which some may find irritating.  It needs to do this for the selftest to work (and for the gdb hooks to be usable if you're actually using them to debug your build of Python).

Should I write a gdb configuration test to check for the availability of gdb built with python?

I've added the file to .hgignore and .bzrignore.  IIRC, a similar thing can be done to the SVN metadata (I don't think this is expressable as a patch, though).  Alternatively, I could wire up the gdb tests to load the file from its location in the source tree.

However, I intend for this code to be installed to a location alongside the build Python, so that it can be automatically detected and used by gdb.  Typically this means copying it to the path of the ELF file with a file.  In my RPM builds I add an extra copy, locating it relative to the location of the stripped DWARF data (e.g. /usr/lib/debug/usr/lib64/

= Selftests =
The selftest runs whatever version of "gdb" is in the path, which then invokes the built version of python, running simple "print" commands and verifying that gdb is corrrectly representing the results in backtraces (even in the face of corrupt data).  I haven't fully tested the error cases yet (e.g. for when gdb is not installed).

The tests take about 14 seconds to run on my box:
[david@brick trunk-gdb]$ time ./python Lib/test/ -v -s test_gdb
The CWD is now /tmp/test_python_19369
test_NULL_ob_type (test.test_gdb.DebuggerTests) ... ok
test_NULL_ptr (test.test_gdb.DebuggerTests)
Ensure that a NULL PyObject* is handled gracefully ... ok
test_classic_class (test.test_gdb.DebuggerTests) ... ok
test_corrupt_ob_type (test.test_gdb.DebuggerTests) ... ok
test_corrupt_tp_flags (test.test_gdb.DebuggerTests) ... ok
test_corrupt_tp_name (test.test_gdb.DebuggerTests) ... ok
test_dicts (test.test_gdb.DebuggerTests) ... ok
test_getting_backtrace (test.test_gdb.DebuggerTests) ... ok
test_int (test.test_gdb.DebuggerTests) ... ok
test_lists (test.test_gdb.DebuggerTests) ... ok
test_long (test.test_gdb.DebuggerTests) ... ok
test_modern_class (test.test_gdb.DebuggerTests) ... ok
test_singletons (test.test_gdb.DebuggerTests) ... ok
test_strings (test.test_gdb.DebuggerTests) ... ok
test_subclassing_list (test.test_gdb.DebuggerTests) ... ok
test_subclassing_tuple (test.test_gdb.DebuggerTests)
This should exercise the negative tp_dictoffset code in the ... ok
test_tuples (test.test_gdb.DebuggerTests) ... ok
test_unicode (test.test_gdb.DebuggerTests) ... ok

Ran 18 tests in 13.233s

1 test OK.
[36833 refs]

real	0m13.599s
user	0m11.771s
sys	0m1.384s

= Platform support =
I don't have access to anything other than Linux, so I've no idea how well this stuff works on other platforms.  My testing so far has been on Fedora, though I've heard of successful usage of this on Debian.

= Legal stuff =

Earlier versions of this code were licensed under the LGPL 2.1
I'm relicensing the code to be "under the same license as Python itself", assuming that's legally OK.  Do I need to state that in the file header, or is that redundant?

I'm in the process of doing the PSF Contributor Agreement paperwork (as an individual); waiting to get my hands on a fax machine.  My employer, Red Hat, has agreed for me to retain copyright on all contributions I make to Python.
msg100491 - (view) Author: Dave Malcolm (dmalcolm) (Python committer) Date: 2010-03-05 16:44
(I faxed in my contributor agreement to the PSF on 2010-03-03)
msg100537 - (view) Author: Martin v. Löwis (loewis) * (Python committer) Date: 2010-03-06 17:28
Maybe I'm using it incorrectly, but my first attempt to use it failed:

gdb) b PyEval_EvalCodeEx
Breakpoint 1 at 0x80fc51d: file Python/ceval.c, line 3020.
(gdb) c

Breakpoint 1, PyEval_EvalCodeEx (co=0xb7d86928, globals=Traceback (most recent call last):
  File "/home/martin/work/27/", line 574, in to_string
    return stringify(proxyval)
  File "/home/martin/work/27/", line 526, in stringify
    return repr(val)
  File "/home/martin/work/27/", line 271, in __repr__
    for arg, val in self.attrdict.iteritems()])
AttributeError: 'FakeRepr' object has no attribute 'iteritems'
, locals=<unknown at remote 0x0>, args=0xb7d7eb00, argcount=1, kws=0x0, kwcount=0, defs=0x827c9b0, defcount=1, closure=<unknown at remote 0x0>)
    at Python/ceval.c:3020
3020            register PyObject *retval = NULL;

This was with the Python 2.7 trunk.
msg100538 - (view) Author: Martin v. Löwis (loewis) * (Python committer) Date: 2010-03-06 17:32
Attribution question: would it be ok if all mentioning of your name in the patch was removed accept for the one in Misc/ACKS? I'm concerned that statements of authorship may become incorrect over time (if people start contributing to it), yet nobody will dare to remove your copyright notice (for example).
msg100662 - (view) Author: Dave Malcolm (dmalcolm) (Python committer) Date: 2010-03-08 20:20
Martin: thanks for reviewing this.

Re msg100537: sorry about the inauspicious start.  I've added some bulletproofing for the case you discovered, and added two new unit tests.

Re msg100538: OK.  I've removed my name and the copyright notice in all places apart from the Misc/ACKS file.  There are still some comments in the code where I use the first person singular.

I'm attaching an updated patch against trunk
msg100882 - (view) Author: Martin v. Löwis (loewis) * (Python committer) Date: 2010-03-11 19:08
I find the printing of frame objects confusing:

3033            f = PyFrame_New(tstate, co, globals, locals);
3034            if (f == NULL)
(gdb) p f
$3 = File /home/martin/work/27/Lib/encodings/, line 15, in decode ()

I didn't recognize that this is actually the output of the Frame; I recommend something like

  Frame %x, for file ....

Also, it prints NULL PyObject* as "<unknown at remote 0x0>". I think null pointers should be special cased, and just be printed as 0x0. Also, what is the "remote" keyword? Aren't all pointers remote in this application? I'd hope that local Python objects never show up.
msg100885 - (view) Author: Dave Malcolm (dmalcolm) (Python committer) Date: 2010-03-11 21:11
Thanks for reviewing the patch.

I've changed the pretty-printing of NULL pointers to "0x0" as suggested, and I've updated frame printing.  Frames are now printed like this
(gdb) p f
$1 = Frame 0x827e544, for file /home/david/coding/python-gdb/, line 19, in <module> () that in a gdb backtrace you might see:
#18 0x080f9aab in PyEval_EvalFrameEx (f=Frame 0x827e544, for file /home/david/coding/python-gdb/, line 19, in <module> (), throwflag=0) at Python/ceval.c:2665

Is this what you had in mind?

As for the use of the word "remote",this is mostly for my own sanity, but I suspect it is needed.  My rationale for this is that there are two address spaces at play: the local address space with the gdb process, and the remote one within the inferior process.  It's possible for addresses with the gdb process to be printed for any object with a NULL tp_repr in its class, and if it were, it would be confusing which address space is it.

For example, if within gdb I run this command:
(gdb) python print repr(gdb.selected_frame())
<gdb.Frame object at 0xb753ef98>

the hexadecimal value that's printed is an address within gdb's address space, that of the python object wrapping the gdb frame information.

I did deliberately model the FakeRepr representation on the output of PyObject_Repr when tp_repr is NULL:
   PyString_FromFormat("<%s object at %p>", Py_TYPE(v)->tp_name, v);
adding the "remote" word to disambiguate things.

Is that OK?  It seemed to me like the best compromise of readability and lack of ambiguity.

I'm attaching an updated version of the patch (version 3), and am about to attach a diff against the last version (which was version 2)

There are some other substantial changes in the patch:
- Line numbers reported in frames were incorrect; I've fixed this, mimicking the "dis" module/PyCode_Addr2Line
- I've started to add various "py-" commands (with selftests):
  - "py-list" command, which tries to mimic gdb's "list" command, but at the level of Python source code
  - "py-up" and "py-down" command, which go up and down the stack, based on the location of PyEval_EvalFrameEx frames
  - "py-bt" which prints a python-level backtrace, based on the location of PyEval_EvalFrameEx frames
- I added a trivial script (Lib/test/ for use by the test cases for the above
- Fixed a bug in PyStringObjectPtr.proxyval() so that it now supports strings with embedded NUL characters
- Various comment fixes.

I split out the test cases into multiple classes, invoking them all.  All pass on my system, taking under 20 seconds ("Ran 28 tests in 17.770s")
msg100886 - (view) Author: Dave Malcolm (dmalcolm) (Python committer) Date: 2010-03-11 21:12
Attaching diff from v2 to v3
msg101720 - (view) Author: Dave Malcolm (dmalcolm) (Python committer) Date: 2010-03-25 16:53
I'm attaching a new version of the patch (v4), against svn trunk (r79422)

Changes since v3:
  * added support for PySetObject (set/frozenset)
  * added support for PyBaseExceptionObject (BaseException)
  * fixed a signed vs unsigned char issue that led to exceptions in gdb for PyStringObject instances containing bytes in the range 0x80-0xff
  * handle the case of loops in the object reference graph.  In previous versions, the gdb prettyprinter would get stuck in infinite recursion following the pointers.  The new version keeps track of the set of all pointers we've visited so far, stopping the traversal at object's we've already reached (analagous to Py_ReprEnter and Py_ReprLeave).
  * unit tests for all of the above (full test suite passes on my machine thus: "Ran 35 tests in 13.547s"; this is a Fedora 12 x86_64 box with gdb-7.0.1-33.fc12)
msg101721 - (view) Author: Dave Malcolm (dmalcolm) (Python committer) Date: 2010-03-25 16:55
(adding diff from v3 to v4, for ease of review)
msg102058 - (view) Author: Dave Malcolm (dmalcolm) (Python committer) Date: 2010-04-01 02:34
I'm attaching a new version of the patch (v5), against svn trunk (r79517)

I've been testing these hooks by using "gdb attach" to attach to real-world python programs.

When doing this with earlier versions of the hooks, gdb would pause for many seconds when printing some values (e.g. during a backtrace).  The issue was that "proxyval" scrapes the entire graph of python objects from the inferior process into gdb, and then prints it.  As well as being slow, this can be many pages of textual output.

To counter this, I've rewritten the pretty-printers to use a stream-based approach.  The various PyObjectPtr subclasses have a "write_repr" method that writes the represenation to a file-like object whilst they recurse.  I supply a file-like object "TruncatedStringIO", which raises an exception after a limit is reached: by default 1K of string data.  (some simple subclasses simply reuse repr(self.proxyval()) for this, but objects that can follow cross-references have their own write_repr implementation).

I've also rewritten frame handling: I eliminated the FrameInfo class, moving its functionality to the PyFrameObjectPtr class, and introduced a new "Frame" class to wrap gdb.Frame.  This allowed a big simplification of the extension commands, and fixed various bugs (involving inlining and optimization)

Within the py-up/py-down and py-bt commands, the code now prefixes Python frame printing with C-level frame indexes, to better tie together the C-level and Python-level views of the stack.

I've added a couple of new commands:
  * "py-print" - takes a string argument.  It attempts to look up the Python variable with that name, and print its value  (it searches first in the locals of the current frame, then globals, then builtins).
  * "py-locals" - prints Python local variables; an analog of gdb's "info locals"

Other changes since v4:
  * Add a ">" marker to the output of "py-list", indicating the current location
  * Add pretty-printer for PyCFunctionObject ("built-in function" and "built-in method")
  * Add selftests for the above
  * Make frame printing fetch information on locals on demand, rather than within FrameInfo's __init__ method (the latter is completely gone)
  * Introduce PyObjectPtr.pyop_field method, simplifying the code
  * Change the set/frozenset selftests to remove a reliance on ordering within the representation of the members
  * Try to be more robust for the case where PyEval_EvalFrameEx's "f" parameter is optimized out (see )
  * Remove the FIXME in PyLongObjectPtr: I've tested the code on builds with both sizeof(digit) == 2 and == 4
  * More selftests for unicode printing

All tests pass on my box (Fedora 12 x86_64 with gdb-7.0.1-33.fc12.x86_64): "Ran 45 tests in 17.439s"
msg102059 - (view) Author: Dave Malcolm (dmalcolm) (Python committer) Date: 2010-04-01 02:37
Adding diff from v4 to v5, for ease of review.

For my reference, md5sum of v5's hooks:
msg102067 - (view) Author: Martin v. Löwis (loewis) * (Python committer) Date: 2010-04-01 07:44
Thanks for the patch. I have committed it as r79548, with two modifications:
- I added a future import for with_statement, as many installations will still use Python 2.5
- I commented out many of the tests, as they were failing for me. I'll report those as several issues.

Please understand that this commit went to the trunk, which will be 2.7. After 2.7b1, no new features are acceptable, so you can only perform bug fixes on the code that is now committed. Please create new issues for any such bug fixes.

Once the tests all pass, I'll merge this into 3.2.
msg102190 - (view) Author: Mark Dickinson (mark.dickinson) * (Python committer) Date: 2010-04-02 19:54
A nitpick:  on OS X, the gdb script ends up being called:

Is this intentional?  If it is, then I'll add this filename to the svn:ignore property.  (And also to make distclean, I guess.  Is currently deleted by a 'make distclean'?)
msg102193 - (view) Author: Dave Malcolm (dmalcolm) (Python committer) Date: 2010-04-02 20:15
> A nitpick:  on OS X, the gdb script ends up being called:
> Is this intentional?  If it is, then I'll add this filename to the 
> svn:ignore property.  (And also to make distclean, I guess.  Is python-
> currently deleted by a 'make distclean'?)

That was unexpected, but re-reading I see where it's coming from.  Please do add it to svn:ignore.

Looks like I forgot to add the removal of the file to the "distclean" target.  Sorry about that.  Presumably adding:
  -rm -f python*
should do the trick (as I understand it, if the glob fails to match, it will happily fail to "rm" the non-existant file named "python*").

Another nit I spotted: the buildbot is "unexpectedly" skipping test_gdb on Win32 (with "gdb not found in path").  Looking at _expectations in Lib/test/ it strikes me that test_gdb is likely to be skipped on every configuration other than on "linux2") - should I wire that in with a decorator within

To my knowledge, OS X doesn't ship with gdb 7 - though presumably someone could build their own copy, linking with the system Python, and if so, the file would then be of use in debugging the freshly-built ./python.exe
msg102195 - (view) Author: Mark Dickinson (mark.dickinson) * (Python committer) Date: 2010-04-02 20:24
Added to svn:ignore in r79616.
msg102197 - (view) Author: Mark Dickinson (mark.dickinson) * (Python committer) Date: 2010-04-02 20:34
> To my knowledge, OS X doesn't ship with gdb 7

That sounds right.  On my OS X 10.6.3 machine, the system gdb is:

Mark-Dickinsons-MacBook-Pro:py3k dickinsm$ gdb --version
GNU gdb 6.3.50-20050815 (Apple version gdb-1346) (Fri Sep 18 20:40:51 UTC 2009)

I have the impression that few people have got gdb 7 working on OS X yet, but I might be wrong.  Certainly macports doesn't seem to offer it yet.  I expect that'll change, though.

[About make distclean]
> Presumably adding:
>   -rm -f python*
> should do the trick

Yes, that seems to work.  Thanks!  Added in r79617.
msg102283 - (view) Author: Benjamin Peterson (benjamin.peterson) * (Python committer) Date: 2010-04-03 17:08
Could I request that the script not be put in the top level directory? It's a little annoying to have another executable script starting with "python" when you're trying to test the "python" executable.
msg102284 - (view) Author: Martin v. Löwis (loewis) * (Python committer) Date: 2010-04-03 17:19
That is tricky to do. gdb will consider it automatically only if it is called <exename>

I agree that there should be a better solution, though. David, is it possible to somehow hook this into .gdbinit, with an arbitrary path?

Benjamin, would a .gdbinit in the top-level build directory still be annoying?
msg102285 - (view) Author: Jeffrey Yasskin (jyasskin) * (Python committer) Date: 2010-04-03 17:27 says "If this file exists and is readable, gdb will evaluate it as a Python script." So maybe it doesn't need to be executable?
msg102286 - (view) Author: Benjamin Peterson (benjamin.peterson) * (Python committer) Date: 2010-04-03 17:31
2010/4/3 Martin v. Löwis <>:
> Benjamin, would a .gdbinit in the top-level build directory still be annoying?

That would be fine with me.
msg102288 - (view) Author: Martin v. Löwis (loewis) * (Python committer) Date: 2010-04-03 18:58
I have now changed to not install as a script anymore; this seems to work fine. People will still need to remove there existing (or make clean) to see this change.
Date User Action Args
2022-04-11 14:56:58adminsetgithub: 52280
2014-02-14 22:52:30jceasetnosy: + jcea
2010-04-03 18:58:23loewissetmessages: + msg102288
2010-04-03 17:31:44benjamin.petersonsetmessages: + msg102286
2010-04-03 17:27:12jyasskinsetmessages: + msg102285
2010-04-03 17:19:52loewissetmessages: + msg102284
2010-04-03 17:08:48benjamin.petersonsetnosy: + benjamin.peterson
messages: + msg102283
2010-04-02 20:34:53mark.dickinsonsetmessages: + msg102197
2010-04-02 20:24:57mark.dickinsonsetmessages: + msg102195
2010-04-02 20:15:44dmalcolmsetmessages: + msg102193
2010-04-02 19:54:56mark.dickinsonsetnosy: + mark.dickinson
messages: + msg102190
2010-04-01 21:36:53dokosetnosy: + doko
2010-04-01 07:44:22loewissetstatus: open -> closed
resolution: accepted
2010-04-01 07:44:13loewissetmessages: + msg102067
2010-04-01 02:37:26dmalcolmsetfiles: + diff-of-gdb7-hooks-v5-relative-to-v4.diff

messages: + msg102059
2010-04-01 02:35:03dmalcolmsetfiles: + add-gdb7-python-hooks-to-trunk-v5.patch

messages: + msg102058
2010-03-28 01:03:46scott.tsaisetnosy: + scott.tsai
2010-03-25 17:28:18jyasskinsetnosy: + jyasskin
2010-03-25 16:55:32dmalcolmsetfiles: + diff-of-gdb7-hooks-v4-relative-to-v3.diff

messages: + msg101721
2010-03-25 16:53:54dmalcolmsetfiles: + add-gdb7-python-hooks-to-trunk-v4.patch

messages: + msg101720
2010-03-25 13:56:29davidfrasersetnosy: + davidfraser
2010-03-11 21:12:51dmalcolmsetfiles: + diff-of-gdb7-hooks-v3-relative-to-v2.diff

messages: + msg100886
2010-03-11 21:11:59dmalcolmsetfiles: + add-gdb7-python-hooks-to-trunk-v3.patch

messages: + msg100885
2010-03-11 19:08:03loewissetmessages: + msg100882
2010-03-08 20:23:02dmalcolmsetfiles: + diff-of-gdb7-hooks-v2-relative-to-v1.diff
2010-03-08 20:20:24dmalcolmsetfiles: + add-gdb7-python-hooks-to-trunk-v2.patch

messages: + msg100662
2010-03-06 17:32:16loewissetmessages: + msg100538
2010-03-06 17:28:38loewissetmessages: + msg100537
2010-03-05 16:44:30dmalcolmsetmessages: + msg100491
2010-03-01 05:01:02gregory.p.smithsetpriority: normal
nosy: + gregory.p.smith
2010-02-28 23:51:53benjamin.petersonsetassignee: loewis

nosy: + loewis
2010-02-28 23:43:21dmalcolmsetfiles: + add-gdb7-python-hooks-to-trunk.patch
keywords: + patch
2010-02-28 23:40:31dmalcolmcreate